Inspectable
How the model works
CORNER does not start from a betting line. Win probability is a logistic transform of a log-odds sum that begins at 50/50. Every term is listed on the matchup page as a factor that moved the projection. If a number is large, you can read why.
Backbone
Each fighter carries opponent-adjusted striking and grappling ratings derived from output, accuracy, takedown volume/defense, control, knockdown rate, and finish mix. Those ratings are weighted by where the fight is likely to be contested — estimated from takedown threat versus takedown defense — so a wrestler’s grappling edge is not applied to a fight that will never hit the mat.
Style interaction
Pressure wrestling against poor takedown defense is not the same fight as pressure wrestling against Merab. Counter strikers get paid against pressure boxers who do not change level. Reach is counted only when the longer fighter actually lives at range (kickers, point fighters, southpaw kick games). Power versus a bad chin is a real term. None of these are allowed to swamp the skill backbone — they are modifiers with conservative priors.
Recency, layoff, camp, age
Last-five results are opponent-quality weighted. A win over a champion is not a win over a journeyman. Single-fight noise is deliberately underweighted. Layoff has a gentle rust prior outside a 60–150 day window. Camp changes are a small negative until there is a sample in the new room. Age decline starts around 30 (32 at heavyweight) and is never treated as destiny.
Five rounds vs three
Championship and main-event time uses cardio ratings and late-round output differential. Grinders and late finishers get paid; fast starters who fade get taxed. Method and expected duration come from the same pieces: knockdown rate, submission rate, estimated standing time, and durability.
Promotions
UFC has the deepest public striking/grappling metrics. PFL regular-season and tournament cards can force finish-seeking; scoring also tends to reward visible pressure. ONE’s smaller gloves, clinch knees, and Muay Thai-literate officiating change the value of kicking and dirty boxing. Thin public data widens the interval. The app will not pretend a Friday Fights strawweight has UFC-grade tracking.
Uncertainty
The point estimate is the center of a band, not a prophecy. 90% intervals come from a standard error that grows with thin data, volatile styles, and large stacked modifiers. CORNER will not show a 92% favorite without showing the interval and the factors. MMA variance is real. The product is the reasoning.
The betting layer
DraftKings and Fanatics are a second, independent data layer. Implied probabilities are vig-removed so you can compare apples to apples. A value flag is model minus market above a threshold. The line is never an input. Fanatics and DraftKings will not always post the same PFL or ONE markets, and they will not always move at the same time. That is displayed, not papered over.
Parlays and pairings
A pairing is two independent sides multiplied. A parlay is three or more. Combined probability is the product of CORNER sides, not the sportsbook parlay screen. Fair price comes from that product; DraftKings parlay price is the product of decimal odds, so juice compounds. Same-card outcomes can correlate through judging, cage, and pace — that covariance is not modeled. A positive EV is a disagreement with the compounded line, not a ticket you are told to place.
Snapshot: fighter stats 19 Aug 2026 (post-UFC 330). Odds several times per day for upcoming cards; this build carries a 19 Aug 2026, 4:00 PM ET book snap.